Aerospace

Wexa AI

Honeywell Aerospace × Wexa AI

AeroFlow

Agentic Supplier Recovery Cockpit

Tool-using agents detect at-risk purchase orders, quantify the revenue they gate through the bill of materials, and draft the recovery — ranked by revenue leverage, not PO count. Every commitment still belongs to a person.

Enter the cockpit →

~3,000

suppliers in the aerospace base

~2%

of them critical or constrained

~$15M

past due at one supplier can gate hundreds of millions

Days → minutes

from signal to an approval-ready recovery plan

The problem

$300M

reduction in midpoint pro forma adjusted EBIT guidance for FY2026, after mechanical supply constraints held output growth below plan.

A handful of suppliers decide whether the quarter lands.

3% → 4%Q1 and Q2 2026 output growth — both below plan, both constrained by mechanical supply rather than demand.
7–9% → 4–5%FY2026 organic growth guidance, cut as a direct consequence.
2 of 3,000percent of suppliers that are critical or constrained. The exposure is concentrated; the attention is not.

Why it is missed today

The signal exists. The number that makes it actionable does not.

Nobody is short of alerts. What is missing is the one figure that turns a late part into a decision: how much revenue it gates, and whose desk that lands on.

Ranked by PO count

Buyers and planners work the queue by volume and by whoever escalated loudest — not by the revenue each line actually holds up.

The answer is in five systems

PO history in SAP, structure in PLM, acknowledgements in the supplier portal, escapes in quality, tooling status in somebody’s inbox. Assembling it is the job.

Where-used is a traversal

Walking a delayed part up the BOM to the end items it gates is a graph problem. A spreadsheet cannot follow it, so in practice nobody does — they estimate.

Days, not minutes

The lag between a slip appearing in the data and a recovery action being taken is measured in days to weeks — most of it spent gathering, not deciding.

The buyer who calls the right supplier first is the one who saw the number first.

The solution

Agents do the assembly. People do the judgement.

AeroFlow is an internal cockpit, not a chatbot. It continuously watches PO, delivery and BOM data, and turns a slipped critical-supplier line into a ranked, priced, citation-backed recovery card within one refresh cycle.

DETECT

Continuous, not on request

Slipped POs from critical or constrained suppliers surface on their own, with the program, the gated end items and the dependency chain attached.

QUANTIFY

Revenue, through the BOM

A where-used traversal converts a late part into dollars gated, with the arithmetic shown line by line so a buyer can check it before acting on it.

DRAFT

Recovery, ready to edit

A root-cause hypothesis, the recovery actions, alternate sources and a supplier email — each claim cited to the source record it came from.

APPROVE

Nothing leaves without a person

No external communication and no system write-back happens without an explicit, logged human decision. Those decisions become the evaluation labels.

How it works

From a late delivery to an approved recovery.

Eight stages, every one of them recorded. Stage seven is a person, and the pipeline cannot route around them.

Signal

A critical-supplier PO slips in the delivery feed

Context

PO history, ack status, inventory, tooling, past-due

Impact

BOM where-used → revenue gated, per part

Risk

Chronic or one-off, with contributing factors

Draft

Recovery playbook and supplier email, cited

Critic

Citations, schema and export policy validated

Human

Buyer approves, edits or rejects

Audit

Immutable run trace, exportable

Long agent work never sits in a request path. Runs are started and polled, so a slow traversal degrades into a slower answer rather than a timeout.

The roster

Bounded autonomy, one responsibility each.

Agent	Responsibility
Orchestrator	Decomposes the goal, routes sub-agents, holds state, enforces escalation thresholds and pauses for human interrupts
Retrieval	Typed tool calls into SAP, PLM, supplier portals, external risk feeds and the vector index — never free-form scraping
Impact	BOM where-used traversal, revenue gated per delayed part, ranked into a leverage list
Risk scoring	Chronic versus one-off supplier risk, with the contributing factors named and a confidence attached
Recommendation	Drafts the recovery playbook and the supplier communication. Never commits anything
Guardrail / critic	Validates citations, schema and export-control policy before a single figure reaches a screen
Human approval	A buyer or planner approves, edits or rejects. This layer is not optional and not automatable

The platform underneath

Built on Wexa Fabric.

Postgres stays the typed system of record. The Fabric context graph is a projection of it — because BOM where-used is a traversal, and a traversal is not a join.

CONTEXT GRAPH

Query live, don’t copy

An ontology-backed graph agents query at run time, so the where-used chain is walked against current data rather than a nightly extract.

APPROVAL GATE

A pause the app cannot skip

Approval is a platform primitive with single-use resume tokens, not an if-statement in application code that a later refactor can drop.

AUDIT TRAIL

Cannot be written, so cannot be forged

A ten-stage lifecycle record per call, written by the platform. AeroFlow has no API to alter it — which is the point.

Trust & governance

Built for a room where “the model said so” is not an answer.

ITARUS-person-only access to ITAR and EAR-controlled records, enforced at the middleware as an attribute on the identity, not as a UI convention.
Gov cloudModel inference behind a gov-cloud boundary with zero retention and no training on inputs, reached through a swappable gateway.
CitationsMandatory source citation plus a critic agent that rejects an unsupported claim before it is ever rendered.
AuditEvery run traced — inputs, tools, model, outputs and the human decision — immutable and exportable for NIST 800-171 / CMMC evidence.

⚠This proof of concept runs on synthetic data only. It is deployed on commercial multi-tenant Fabric (fabric.wexa.ai), which is not a gov-cloud boundary. No real Honeywell PO, BOM or supplier data may be loaded until that decision is resolved — and a CI guard fails the build on a real-looking identifier.

What it has to move

Pilot targets, agreed before the build.

Leading indicators prove the loop is working; lagging ones prove it mattered. Baselines are measured in Sprint 0, so none of these is graded against a memory.

Measure	Type	Pilot target
Time to detect an at-risk PO	Leading	−50%
Time to first action	Leading	−40%
Past-due revenue-at-risk with an active recovery plan	Leading	≥80%
Buyer hours saved per week, per FTE	Leading	≥6
Past-due backlog on the pilot family	Lagging	−15%
On-time delivery on the pilot family	Lagging	+3–5 pts
Recommendation acceptance rate	Quality	≥70%
Hallucination rate on the eval set	Quality	<2%

Ready when you are

Ranked by revenue leverage. Approved by people.

Aerospace

Wexa AI

AeroFlow — internal proof of concept. Synthetic data only.Cockpit System health